OptFin.org | Public Learning Studio

Public academic demonstration

Dimensional Profile Permanent Collapse Spectral Islands Asynchronous Swarm RL Thermostat

Control Room

Profile calibration

Tune the instance scale and the collapse rules. The simulator recalculates the active volume, CPU effort, and convergence path continuously.

Variables (N) 5000

Tasa Apr. 0.42

Freeze threshold 99.2%

Collapse pressure 0.63

Islas CPU 4

Swarm aggressiveness 0.58

Interpretation

What the machine is doing

The engine is profiling the instance to infer safe subgraph sizes, learning rate, and how quickly the tensor may freeze a variable without blinding the search.

Volumen N

Active domain collapse

Each particle represents a slice of the active decision graph. Bright particles remain unresolved. Dim particles have been permanently extracted into the frozen core and no longer consume active solver memory.

Progreso Fitness

Runtime telemetry

Iter. 0

Var. N 5000

Tasa Apr. 0.42

Islas CPU 4

Coste CPU 1500 ms

Motor Profiling

Variables (N)

CPU / Iteration

Fitness

Stage 1 Profile

Stage 2 Extract

Stage 3 Partition

Stage 4 Swarm

Stage 5 Thermostat

Engine Trace

Permanent collapse timeline

The timeline below shows how certainty harvest reduces the active search frontier. As the core vanishes, the CPU cost per iteration falls and the swarm can spend more of the budget on difficult residual neighborhoods.

Active frontier preserved for exact refinement Frozen core 0 variables

Dimensional Profile

Auto-calibration before search

The profile estimates how many variables can be treated safely in each exact kernel, how many spectral islands should be opened, and how patient the thermostat must be before it reheats the search.

Permanent Collapse

Certainty becomes structure

Frozen variables are not only fixed. They are removed from the active instance. Their profit or cost is harvested, capacities are updated, and the residual problem becomes physically smaller.

Asynchronous Swarm

Parallel exactness without idle cores

Once the graph has been partitioned, each island can carry its own exact or bounded search lane while the central tensor keeps learning from the best structures found so far.